FGA-Net: Fourier-guided attention network for feature enhancement in colorectal cancer digital pathology images

Accurate classification of colorectal histopathology images is challenging because diagnostically relevant patterns are distributed across both fine-grained cellular textures and broader tissue architecture. Conventional handcrafted descriptors often lack sufficient discriminative capacity, while purely spatial deep models do not explicitly exploit frequency-domain information.

We propose FGA-Net, a Fourier-Guided Attention Network for feature enhancement in colorectal cancer digital pathology images. The model combines patch embedding, transformer-style encoding, and a Fourier-guided Attention for feature Refinement (FAR) module integrating FFT/IFFT transformation, squeeze-and-excitation recalibration, and self-attention. Experiments were conducted on the public EBHI dataset under the binary benign-versus-malignant setting at 200× magnification.

FGA-Net consistently outperformed five handcrafted feature families across seven downstream classifiers. The best configuration, FGA-Net with ANN, achieved an accuracy of 87.54%, surpassing the strongest handcrafted baseline by 11.52 percentage points. Ablation studies further showed that both the spectral branch and the channel recalibration branch contributed positively to performance.

The results indicate that Fourier-guided attention effectively improves pathology feature representation by jointly modeling spectral cues and long-range spatial dependencies. FGA-Net provides a more discriminative and transferable representation for colorectal histopathology classification and offers a promising direction for future computational pathology research.